4 ms·
This suffers from the problem of sparse rewards. You need a better reward that gives value to intermediate results or special-case answers, since too few states
by jhanschoo 3y ago
This suffers from the problem of sparse rewards. You need a better reward that gives value to intermediate results or special-case answers, since too few states are correct.
- t_mann 3y agoIt's not about training the model. You just input the error messages together with an appropriate prompt and ask it to fix it. Iterate until it verifies successfully (or stop after you've reached ~1-2x content window length). Once you have produced examples of correct math that way, you can use those to fine-tune the model.